[PATCH 7/7] boot: add a minimal bootmeth for the Boot Loader Specification

[Simon Glass]

Hi Alexey,

On Thu, 25 Jun 2026 at 17:57, Alexey Charkov <alchark at flipper.net> wrote:
>
> On Thu, Jun 25, 2026 at 8:24 PM Simon Glass <sjg at chromium.org> wrote:
> >
> > Hi Alexey,
> >
> > On Thu, 25 Jun 2026 at 16:52, Alexey Charkov <alchark at flipper.net> wrote:
> > >
> > > Hi Simon,
> > >
> > > On Thu, Jun 25, 2026 at 7:28 PM Simon Glass <sjg at chromium.org> wrote:
> > > >
> > > > Hi Alexey,
> > > >
> > > > On 2026-06-04T15:31:05, Alexey Charkov <alchark at flipper.net> wrote:
> > > > > boot: add a minimal bootmeth for the Boot Loader Specification
> > > > >
> > > > > Add a bootmeth that finds and boots Boot Loader Specification (BLS)
> > > > > type #1 entry files [1]. On each block-device partition it scans, the
> > > > > bootmeth looks for files matching '<prefix>loader/entries/*.conf'
> > > > > (where <prefix> comes from bootstd_get_prefixes(), typically '/' and
> > > > > '/boot/'), picks the highest-sorting filename, parses it, and exposes
> > > > > it as a bootflow.
> > > > >
> > > > > Implementation reuses the existing pxelinux infrastructure.
> > > > >
> > > > > For now the entry chosen on a partition is purely the lexicographic
> > > > > maximum of *.conf filenames; sort-key / version field handling
> > > > > (spec-mandated tiebreakers) and boot-counting (the '+TRIES_LEFT'
> > > > > filename suffix) are left as TODOs. Likewise, only the top-sorted entry
> > > > > is surfaced because the bootstd framework currently allows one bootflow
> > > > > per (bootmeth, partition); exposing every discovered entry will require
> > > > > a framework extension.
> > > > >
> > > > > Type #2 BLS (drop-in directory of EFI binaries) is out of scope here;
> > > > > [...]
> > > > >
> > > > > boot/Kconfig        |  17 +++
> > > > >  boot/Makefile       |   1 +
> > > > >  boot/bootmeth_bls.c | 333 ++++++++++++++++++++++++++++++++++++++++++++++++++++
> > > > >  3 files changed, 351 insertions(+)
> > > >
> > > > > diff --git a/boot/Kconfig b/boot/Kconfig
> > > > > @@ -649,6 +649,23 @@ config BOOTMETH_EXTLINUX_PXE
> > > > > +config BOOTMETH_BLS
> > > > > +     bool "Bootdev support for Boot Loader Specification entries"
> > > > > +     select PXE_UTILS
> > > > > +     default y
> > > >
> > > > Wherever the 'default y' discussion lands, this will be enabled on
> > > > sandbox, so the 'bootmeth list' tests in test/boot/bootmeth.c will
> > > > fail since they check the exact list and count. Please can you run the
> > > > sandbox tests and update them as needed?
> > > >
> > > > Since you are adding a new bootmeth you also need a sandbox test that
> > > > exercises it (see the extlinux tests in test/boot/bootflow.c, with a
> > > > fixture disk image) and a documentation page, e.g.
> > > > doc/develop/bootstd/bls.rst with an entry in the index there.
> > >
> > > Will do, thanks!
> > >
> > > > > diff --git a/boot/bootmeth_bls.c b/boot/bootmeth_bls.c
> > > > > @@ -0,0 +1,333 @@
> > > > > +static int bls_getfile(struct pxe_context *ctx, const char *file_path,
> > > > > +                    char *file_addr, enum bootflow_img_t type, ulong *sizep)
> > > >
> > > > This is a verbatim copy of extlinux_getfile() - please can you export
> > > > that, or move it into a shared helper?
> > >
> > > Ack
> > >
> > > > > diff --git a/boot/bootmeth_bls.c b/boot/bootmeth_bls.c
> > > > > @@ -0,0 +1,333 @@
> > > > > +     ret = bootmeth_alloc_file(bflow, SZ_64K, 1, BFI_EXTLINUX_CFG);
> > > >
> > > > The extlinux bootmeth passes ARCH_DMA_MINALIGN here, since the buffer
> > > > is filled by a block-device read which may use DMA on some platforms.
> > > > An alignment of 1 risks cache-line corruption there, so please use
> > > > ARCH_DMA_MINALIGN.
> > >
> > > Ack
> > >
> > > > > diff --git a/boot/bootmeth_bls.c b/boot/bootmeth_bls.c
> > > > > @@ -0,0 +1,333 @@
> > > > > +     bflow->bootmeth_priv = label;
> > > > > +     free(fpath);
> > > >
> > > > This leaks memory: bootflow_free() releases bootmeth_priv with a plain
> > > > free(), so the label's string members (name, menu, kernel, append,
> > > > initrd, etc.) are never freed - for every bootflow discarded after a
> > > > scan, not just on error paths.
> > > >
> > > > Since get_string() copies tokens out of the buffer rather than
> > > > modifying it in place, you could follow the extlinux approach: keep
> > > > only bflow->buf across the scan (already populated by
> > > > bootmeth_alloc_file() and freed by the framework), use a temporary
> > > > label in bls_read_bootflow() just to extract the title, destroy it
> > > > with label_destroy(), and re-parse the buffer in bls_boot(). Then
> > > > bootmeth_priv is not needed at all. What do you think?
> > >
> > > Will address the leak, thanks for pointing it out!
> > >
> > > The bls_priv structure is needed for my follow-up extension which
> > > allows multiple boot entries to be returned by each (bootdev,
> > > bootmeth, partition) tuple, and I wanted to minimize churn between
> > > those two. I haven't sent that follow-up for review yet, as I wanted
> > > to confirm this simpler version is acceptable first.
> >
> > We should implement this using an generic index rather than something
> > bls-specific...please see some commits at:
> >
> > https://concept.deinde.dev/u-boot/u-boot/-/merge_requests/782
>
> Oh, there's already BLS in your tree :-D
>
> Would you like to merge that instead? Your version looks much more
> full-fledged. If you happen to have one rebased on master or next I'd
> be happy to test!

Unfortunately there is a lot of pxe refactoring, because I wanted to
make it possible to fully parse an extlinux file (plus includes)
without actually booting anything. I actually did look at it a few
weeks back.

Your approach is easier. I can easily add my patches on top of yours
once they land.

>
> In my attempt at multi-bootflows I just pass an extra 'seq' parameter
> to each bootmeth. If all bootflows on a partition are exhausted, or if
> the bootmeth in question doesn't support multiple bootflows per
> partition (and seq > 0 is passed) they return -ENOENT, so that the
> caller can proceed to the next (bootdev, bootmeth, partition) tuple.
> Otherwise the caller repeats with the same tuple until it gets
> -ENOENT. So no extra flag to indicate the multi-bootflow capability as
> in your version.

I'm not wedded to the flag - we can update other bootmeths to return
-ENOENT if the seq is non-zero.

>
> For BLS, I store an alist caching all discovered entries per partition
> inside the `priv` structure to avoid needlessly repeating directory
> operations.

So long as you handle clearing that out, that seems fine.

>
> > > > > diff --git a/boot/bootmeth_bls.c b/boot/bootmeth_bls.c
> > > > > @@ -0,0 +1,333 @@
> > > > > +     ret = pxe_setup_ctx(&ctx, &cmdtp, bls_getfile, &info, true,
> > > > > +                         NULL, false, false);
> > > >
> > > > Just to check, how does this behave when the entry is found under a
> > > > non-empty prefix? bls_pick_entry() searches every bootstd prefix, but
> > > > the prefix the winning entry came from is not recorded. The spec says
> > > > paths in an entry are relative to the root of $BOOT, so for an entry
> > > > under '/boot/loader/entries/' the kernel must be fetched from
> > > > '/boot/...', yet with allow_abs_path set and a NULL bootfile,
> > > > get_relfile() uses the absolute path from the partition root and the
> > > > lookup will miss. I suspect you need to remember the discovered prefix
> > > > (bflow->subdir is a natural place, and is freed automatically) and
> > > > prepend it when fetching files.
> > >
> > > Hmm, I am booting it with a non-empty prefix and it works as-is (I use
> > > a single ext4 partition for the root fs, and my entries are under
> > > /boot, i.e. there is this "boot/" prefix from the root of the
> > > partition to get to "loader/entries").
> > >
> > > My loader entries use full absolute paths including the "/boot"
> > > component, which is the way systemd's 90-loaderentry.install creates
> > > them in this setup. Or did you mean that this behavior is not
> > > spec-compliant? Not sure if prefixes are spec-compliant anyway, but
> > > definitely useful as an extension.
> >
> > Ah OK. I was thinking of extlinux - but what happens if you don't use
> > full absolute paths? There is the case where the boot partition is a
> > separate filesystem and the OS files are there, so the running system
> > sees the file as e.g. /boot/vmlinux (since the boot partition is
> > mounted at /boot) but U-Boot sees it as /vmlinux (since it does not
> > have a VFS).
>
> I will recheck with a separate partition and paths stemming from its
> root (i.e. /vmlinuz or /[entry-key]/[entry-id]/linux), thank you!

Good to have a test :-)

Regards,
Simon